[fix] Polish code of `models`

[tomiock]

Qwen3-VL and Qwen3.5 share a lot of code, initially file was copy-pasted over.
We can recycle a lot of snippets between the three models. Perhaps even abstracting the visual model definition.

[tomiock]

• Added kernel conv1d from Tao. We rely on a torch.bucketize to prepare the sequence index for the kernel:

 # Projections
       qkv = self.in_proj_qkv(x)  # (B, L, conv_dim) — channel-last in memory
        z = self.in_proj_z(x)
        a = self.in_proj_a(x)
        b = self.in_proj_b(x)

        # Per-token segment index for causal_conv1d_fn's packed mode. Using
        # bucketize keeps this graph-traceable with no host sync.
        seq_idx = torch.bucketize(
            torch.arange(L, device=x.device), cu_seqlens[1:-1], right=True
        ).to(torch.int32).unsqueeze(0).expand(B, -1).contiguous()

        # Fused causal-conv1d + SiLU. Input is (B, C, L) with channel-last
        # strides (stride(1) == 1), which is what the kernel expects when
        # seq_idx is provided.
        mixed_qkv = _causal_conv1d_fn(
            x=qkv.transpose(1, 2),
            weight=self.conv1d.weight.squeeze(1),
            bias=self.conv1d.bias,
            seq_idx=seq_idx,
            activation="silu",
        ).transpose(1, 2)  # (B, L, conv_dim)

[tomiock]

The visual attention now works with attn.varlen from torch. No loss changes.